ANALYSIS OF PRICE DYNAMICS OF BLOCKCHAIN-BASED CARBON CREDIT TOKENS IN THE CRYPTOCURRENCY MARKET USING DEEP LEARNING METHODS

Abstract

The aim of this study is to analyze the price dynamics of blockchain-based carbon credit tokens, namely Base Carbon Tonne (BCT), Moss Carbon Credit (MCO2), and KlimaDAO (KLIMA) as well as mainstream crypto assets such as Bitcoin (BTC), Ethereum (ETH), Cardano (ADA), and Solana (SOL) and the speculative asset Carboncoin (CARBON). In addition, the Fear & Greed Index, which represents investor sentiment, has been incorporated into the model in line with the role of sentiment-driven effects in price formation processes in cryptocurrency markets, as highlighted in the literature. The study utilized daily closing prices from the period October 21, 2021, to November 1, 2025; correlation analyses were performed on raw daily price series using the Pearson correlation method, which was chosen to examine the direction and strength of the linear relationship between variables. Prior to modeling, the dataset was cleaned, Min-Max normalization was applied, and it was split into a 70% training set and a 30% test set while preserving chronological integrity. While the assumption of stationarity in time series is important from the perspective of classical econometric approaches, this study focuses on deep learning-based methods within the scope of nonlinear modeling frameworks. The data used in the study were obtained from Yahoo Finance and the AI Key API. The findings indicate that there are strong internal linkages among carbon credit tokens. In particular, while a strong positive relationship was observed between BCT and MCO2, it was determined that these tokens exhibit a weak negative correlation with Bitcoin. This suggests that carbon credit tokens are only marginally linked to the broader crypto market but form a more cohesive structure within their own ecosystem. Additionally, it was observed that the CARBON asset exhibits relationships ranging from weak to moderate with major crypto assets. The Fear & Greed Index, meanwhile, showed moderate relationships with BTC, ETH, and SOL, and weaker relationships with carbon credit tokens. During the modeling process, LSTM, GRU, Transfer-LSTM, and Transfer-GRU architectures were used; the data was split into 70% training, 30% validation, and 30% test sets while maintaining chronological integrity; the models were evaluated using MSE, RMSE, MAE, MAPE, and R² metrics. The results show that the GRU architecture generally offers the highest prediction accuracy, while transfer learning models perform relatively better in predictions for the KLIMA and Fear & Greed (F&G) Index. Overall, the study demonstrates that deep learning and transfer learning approaches are effective in modeling price behavior in tokenized carbon credit markets. Here, it is assessed that transfer learning does not automatically provide an advantage in every scenario, but offers strategic contributions for specific asset groups. In conclusion, the study demonstrates that AI-based models can be used as a decision-support mechanism in the pricing of sustainable financial instruments in the digital economy.